4DAnyone

Why we included this project

Anyone building virtual humans, avatars, or motion-capture-style pipelines should look at this: 4DAnyone turns a single casual phone-shot video of one person into a dense set of multi-view videos around them, which you then feed into 4D Gaussian splatting to reconstruct a free-viewpoint, animatable human. No rigged capture setup, multi-camera array, or tripod is required, and preset camera layouts range from a compact 6-view orbit up to a 48-view three-pitch ring, depending on how much coverage your reconstruction needs. The repo documents the performance work too, with roughly 25 GiB peak GPU memory, a 1.42x end-to-end speedup on the full 24-view pipeline, and support for attention backends like FlashAttention-3 and SageAttention. Inference is scripted, with plain arguments like view count and yaw span, so it behaves more like a clean tool than a typical research checkout.